I spent some time this year preparing a technical briefing for the Pathogen Genomics Initiative team, built around the 2025 edition of the Front Line Genomics Sequencing Buyer’s Guide — a market snapshot covering extraction, library preparation, short- and long-read platforms, and outsourcing. It is a useful catalogue. It is also, as its own foreword states, sponsored by two companies that sell into the market it surveys. That is not a reason to discard it, but it is a reason to treat its comparison tables as a starting inventory rather than as an evaluation.
Working through roughly forty platforms, the thing that struck me is how little of the purchasing decision the specifications actually determine.
Vendors compete on output per run, read length and quality score. Q30 means an estimated one error per thousand bases; Q40, one per ten thousand. These are estimated per-base error probabilities. They are not the accuracy of a finished genome, and they are certainly not the reliability of a public health interpretation. Coverage depth cannot recover a region the assay never amplified, and it does not correct a systematic error.
What a surveillance programme actually consumes is reportable genomes — samples that pass the complete workflow and produce something a public health decision can rest on. So the procurement metric I proposed to the team was cost per reportable genome: total attributable workflow cost divided by the number of samples meeting reporting criteria. It is a harder number to obtain than a price per gigabase, because no vendor publishes it. That is rather the point.
A worked example, with illustrative figures rather than a recommended threshold. Take a bacterial genome of 5 Mb at 50× depth: 250 Mb of target bases per sample. Take a run producing 15 Gb, of which perhaps 70% survives filtering and is on-target: 10.5 Gb of effective output. That divides into about 42 samples. Ask for 100× instead and it halves to 21.
Then subtract your controls. Then account for uneven pooling, and for the libraries that underperform. The headline capacity on a datasheet and the number of genomes you report at the end of the month are different quantities, and the gap between them is where budgets are lost.
Fixed run costs reward filling the flow cell. A hypothetical run costing US$1,200 works out at US$50 per sample across 24 samples, US$25 across 48, and US$12.50 across 96. The arithmetic is seductive, and it quietly argues for waiting until you have a full batch.
During an outbreak that argument is wrong. Waiting to fill a run converts money saved into days lost, and days are the currency of outbreak response. Any honest operating plan has to model both: a routine mode optimised for cost per result, and a surge mode that accepts a half-empty flow cell.
The same realism applies to turnaround. Instrument run time is one segment of the interval between specimen receipt and an approved report. Transport, extraction, library preparation, queueing, transfer, analysis, QC review and release make up the rest. A fast instrument in a slow system produces slow reports.
Site readiness decides more installations than chemistry does: power stability and backup, environmental conditions, reagent shipping and shelf life, minimum order quantities, replenishment lead times, trained operators, and — the one most often discovered too late — whether the manufacturer or distributor actually provides engineering support in the country of installation, as distinct from the region.
Then there is the data. Versioned analysis pipelines, compute sufficient for timely basecalling and demultiplexing, documented export of reads and metadata, retention and access governance, and interoperability with the repositories a programme is obliged to submit to. Deciding this after the instrument arrives is expensive.
I proposed a weighted scorecard to structure the comparison: assay fit and demonstrated quality 30%, total cost of ownership 25%, supply and service 20%, reporting speed and capacity 15%, data and workforce fit 10%.
The weights matter less than one rule that sits above them. Mandatory requirements are applied before weighted scoring. A vendor that fails a must-have does not get rescued by a low price or a strong total — otherwise the scorecard becomes a mechanism for talking yourself into a bad purchase.
And the scorecard does not end the process. Acceptance criteria should be agreed before delivery, and the pilot should run representative specimens — routine and difficult — with proper controls, measuring QC success, reporting time, concordance and repeat rate on complete reports rather than on vendor run metrics.
The last point is the one most resisted. Sequencing capacity is a network property, not a per-laboratory badge. Predictable recurring workload justifies in-house capacity. Low, intermittent or specialist demand often does not, and a referral or contracted arrangement with clear terms on quality, logistics, data return and raw-data access will serve the public health objective better than an underused instrument with an unfunded service contract.
An instrument that sits idle is not capacity. It is a depreciating asset with a cold chain attached.
The briefing itself was prepared for internal discussion; platform details derive from the Front Line Genomics guide and require configuration-specific confirmation with vendors before any procurement.
]]>Everything is now held in _data/*.yml. A course is a two-line YAML entry. An
award is three. The publication list is not written by hand at all — it is
generated from my OpenAlex record by a script in the
repository, so adding a paper means running one command.
Two deliberate constraints:
_data/pub_overrides.yml, which the generator reads and
never rewrites, so the correction survives every refresh.The source is on GitHub if any of it is useful to you.
]]>